read_phy / PhySortingExtractor raises "assignment destination is read-only" under pandas >= 3.0

[rly]

Description

BasePhyKilosortSortingExtractor.__init__ writes into a numpy array obtained from DataFrame.values. Since pandas 3.0 made Copy-on-Write permanent, .values returns a read-only array, so the in-place assignment raises ValueError: assignment destination is read-only. This makes read_phy fail on any Phy folder that spikeinterface itself exported (i.e. one containing an si_unit_id column).

This was previously noted in passing by @MarinManuel in #4231 (comment); filing it as a dedicated bug with a minimal reproduction and a proposed fix. It is downstream-visible: it breaks neuroconv's PhySortingInterface (and NWB GUIDE's tutorial, NeurodataWithoutBorders/nwb-guide#1080) whenever pandas 3.0 is installed. Note spikeinterface[full] pins pandas<3 and so is unaffected, but the base install is not.

Location

src/spikeinterface/extractors/phykilosortextractors.py in BasePhyKilosortSortingExtractor.__init__ (line numbers from current main):

168:  unit_ids = cluster_info["si_unit_id"].values   # read-only under pandas >= 3.0
...
183:  unit_ids[i] = new_si_id                         # ValueError: assignment destination is read-only

Minimal reproduction

Uses only spikeinterface + its dependencies. Verified on main and on 0.104.8.

import tempfile
from pathlib import Path

import numpy as np
import pandas as pd
import spikeinterface
from spikeinterface.extractors import read_phy

folder = Path(tempfile.mkdtemp()) / "phy"
folder.mkdir()

np.save(folder / "spike_times.npy", np.array([10, 20, 30, 40], dtype="int64"))
np.save(folder / "spike_clusters.npy", np.array([0, 1, 0, 1], dtype="int64"))
(folder / "params.py").write_text("sample_rate = 30000.\n")
# `si_unit_id` is the column spikeinterface writes when it exports a sorting to Phy.
(folder / "cluster_info.tsv").write_text("cluster_id\tsi_unit_id\n0\t0\n1\t1\n")

print("spikeinterface", spikeinterface.__version__, "| pandas", pd.__version__)
read_phy(folder_path=folder)   # ValueError: assignment destination is read-only  (pandas >= 3.0)

• pandas 2.3.3 -> works.
• pandas 3.0.0 -> raises.

Traceback

File ".../spikeinterface/extractors/phykilosortextractors.py", line 183, in __init__
    unit_ids[i] = new_si_id
ValueError: assignment destination is read-only

Suggested fix

Take an owned, writeable copy before mutating:

unit_ids = cluster_info["si_unit_id"].to_numpy(copy=True)

(Worth auditing the other cluster_info[...].values uses in this method for the same pattern.)

Environment

• spikeinterface: main (also reproduced on 0.104.8); the buggy line is identical back to 0.103.0, so it is version-independent
• pandas: 3.0.0
• numpy: 2.x
• Python: 3.12

[alejoe91]

@rly we are a bit behind on pandas 3.0. All our test suite is running on pandas<3...we should update and spot any other issues with that.

[alejoe91]

Let's see how it goes: #4691

[rly]

Thanks!